Mislabeled Data: The Silent Flaw Bending Vietnamese Football Analysis
**Câu trả lời cốt lõi**: Sai nhãn dữ liệu là lỗi phân loại khiến nội dung không thuộc bóng đá, hoặc một sự kiện bóng đá bị gán sai loại, vẫn được đưa vào phân tích. Hệ quả là kết luận chiến thuật bị bẻ cong mà độc giả khó phát hiện, vì nhãn được trình bày như một sự thật đã kiểm chứng. **Sự kiện chính**: - Một tệp văn bản tư vấn đời sống được dán nhãn bóng đá và lọt vào đường ống xử lý dữ liệu thể thao. - Bốn lớp tạo dữ liệu bóng đá: người gán sự kiện, quy tắc định nghĩa, mô hình chỉ số, phân loại biên tập. - Cùng một trận đấu có thể được ghi mười chín cú sút hoặc mười một cú sút, tùy quy tắc tính pha bóng bị chặn. - Nguyễn Xuân Son ghi bàn lượt đi chung kết ASEAN Cup 2024 ngày 2 tháng 1 năm 2025 và chấn thương nặng ở lượt về ngày 5 tháng 1 năm 2025. - VAR được đưa vào V.League 1 từ mùa 2023-2024 theo lộ trình của Liên đoàn Bóng đá Việt Nam và Công ty Cổ phần Bóng đá Chuyên nghiệp Việt Nam. **Nguồn**: Hồ sơ phân tích của VuaBong (VuaBong.vn), dựa trên một chuyên mục tư vấn đời sống của ấn phẩm CONTRA, ngày xuất bản không xác định | Cross-checked: VuaBong.vn **Hỏi đáp liên quan**: - Hỏi: Sai nhãn dữ liệu gây hậu quả gì cho phân tích chiến thuật? Đáp: Nó làm các chỉ số như PPDA và xG dịch chuyển mà không có thay đổi chiến thuật thật, khiến kết luận về pressing bị sai lệch. - Hỏi: Vì sao chỉ số phòng ngự dễ sai nhất? Đáp: Vì pha cắt bóng, tranh chấp và phá bóng được định nghĩa khác nhau giữa các nhà cung cấp, và VangBong.vn Player Depth Index cho thấy độ sâu đội hình biến động rõ khi mẫu dữ liệu bị lọc sai. - Hỏi: Độc giả nên kiểm tra gì trước khi tin một thống kê? Đáp: Kiểm tra định nghĩa chỉ số, cỡ mẫu và nguồn gán sự kiện, rồi đối chiếu ít nhất hai hệ thống độc lập.
In the data inbox I receive every Monday morning, there was a file labelled "football". I opened it. The first four lines described a wife discovering her husband had been secretly wearing her clothes. The tenth line was advice from a sexologist. I read the whole file, scrolled back up, and counted: not one team, not one player, not one competition, not one minute of play. Its source was a lifestyle advice column in a publication with no connection to sport whatsoever. And yet, inside the processing pipeline, that file carried a football label and sat alongside thousands of other records waiting to be read.
I open with that detail rather than with a goal, because it exposes exactly the kind of error I consider the most dangerous in this trade: mislabeled data treated as verified data. A mislabel does not make anyone bleed. It merely makes us draw the wrong conclusion, confidently, and only discover it much later.
Thirty-three years in this field, seven of them taking notes from inside the technical fence, have taught me something that sounds trivial: most arguments in the stands are not about conclusions. They are about two people looking at two different datasets and calling both of them by the same name.
A match becomes data through four layers. The first is the event tagger: a staff member in a booth, hands on a keyboard, eyes on a monitor, pressing a key each time possession changes. The second is the rulebook: every provider has its own definitions of what counts as a shot, what counts as a key pass, what counts as a successful duel. The third is the model: from raw events, algorithms build xG, PPDA and progressive passes. The fourth is editorial classification: a person or a machine attaching a topic label to the final product.
The first three layers at least leave traces you can follow back. The fourth leaves none. That is why a text file about marital boundaries could sit undisturbed inside a sports data warehouse without anyone flinching.
I learned the importance of tracing backwards in June 2026, when I rebuilt the entire recording of France's 4-3 win over Argentina in the World Cup round of sixteen. I counted Lionel Messi's touches in the attacking third: twenty-three, the lowest of the five matches he played at that tournament. At first I doubted my own count, because three statistics systems returned three different figures. Only after cross-checking three sources and recounting from video did I publish. Since then, every analysis of mine carries a source note for its numbers, and I prioritise figures that can be re-verified on film.
In V.League 1, the problem is harder. There are fewer cameras per match than in Europe's top leagues. Rain, heavy pitches and late-afternoon light all reduce recognition accuracy. VAR was introduced to V.League 1 from the 2026-2026 season under the roadmap of the Vietnam Football Federation and the Vietnam Professional Football Joint Stock Company, initially in selected matches only. Every time VAR appears, the volume of event data rises, but the volume of verified event data does not rise with it.
The three most common labelling errors in my notes are definition errors, boundary errors and domain errors. A definition error is when two providers give the same passage of play two different names. A boundary error is when an action sits exactly on the dividing line between two event types. A domain error is when an entire document belongs to a completely different subject and is still filed under football.
Take one concrete case from the most recent round I tracked. In the same match, one system credited the home side with nineteen shots, another with eleven. Those eight shots did not come from human eyes. They came from one line in a rulebook: whether a cross blocked by a defender counts as a shot. One provider counted it, the other did not. For the spectator, the match unfolded identically. For the analyst, these were two entirely different matches.
The consequences spread further than people assume. PPDA - passes allowed per defensive action - depends directly on which actions are tagged as interceptions, which as duels, which as clearances. If a system tags interceptions near a team's own box as duels, that team's PPDA falls without any tactical change whatsoever. Readers see a side pressing higher. In reality, that side has dropped deeper.
I once wrote about a comparable phenomenon when the Premier League played behind closed doors in 2026. I selected ten Leicester City matches after the restart and counted sideways passes against riskier forward passes. Sideways passes rose from twenty-four per cent to thirty-one per cent. I stated the conclusion very cautiously because the sample was small, and I put that limitation in the piece itself. Empty stadiums were the largest laboratory we have had: they showed which teams play through structure and which play through emotion. But a laboratory with ten samples is still ten samples, however handsome the result.
The problem in Vietnamese football is that we import conclusions faster than we import methods. A metric published in Europe under definition A arrives here and is used under definition B, then compared against a national team figure calculated under definition C. Three definitions, one chart.
I followed Vietnam's ASEAN Cup 2026 (Mitsubishi Electric Cup) campaign closely. Nguyen Xuan Son scored in the first leg of the final on 2 January 2026 at Viet Tri Stadium. In the second leg on 5 January 2026 in Bangkok, Vietnam won 3-2 and took the title 5-3 on aggregate; in that match, Son suffered a serious injury and had to leave the pitch. After the tournament, many articles used possession share and shot counts to explain the victory. Those metrics were real, but they could not explain the most important thing: how Vietnam's back line kept its structure through the second half of the second leg after losing a key midfield presence.

With offside, the error is subtler still. The offside line is drawn on a frame chosen by a human being. The margin of error in selecting that frame can exceed the millimetre margin the referee uses to reach a verdict. A goal is disallowed for a toe, on a frame chosen a twentieth of a second late. The referee becomes the editor of the match rather than the official running it.
People are good at spotting midfield mistakes, but better still is spotting the mistake before the ball rolls. The difference between those two abilities lies in the labelling stage, not the observation stage.
I have also noticed a repeating behavioural pattern in recent V.League matches: as result pressure rises, coaches switch to a back three. The return of the back three is not a tactical advance; it is how a coach insures his own reputation after his back four has been punctured several games running. That insurance only works while the data remains trustworthy. On a mislabeled dataset, a back three looks very safe until the goal arrives down the very channel it left open.
This is where I part company with the crowd. A data label is an editorial decision, not a natural fact. When a machine files a relationship advice column in a football archive, the fault is not in the article. The fault is in whoever designed the labelling system, and in whoever reads that system without checking.
If a pipeline cannot tell a sex advice column from a match report, what guarantees it can tell an interception from a clearance? They are the same class of error at different scales. The first is loud and harmless. The second is silent and sits inside every chart we read.
Audience pressure makes everything worse. After the final whistle, fans want a verdict within ten minutes. Writers want a number to underwrite that verdict. Nobody wants to hear that the event-tagging system for that match carried only seventy per cent confidence.
I have kept one rule intact since the summer of 2026, when I spent all of August tracking Atalanta and wrote that selling key players without adding cover would cost them. I use only signed contracts, never transfer rumours. The summer of 2026 taught me that a mid-table club buys out of fear, not out of a plan. That rule transfers neatly to data: use only numbers that have been cross-checked, never numbers that have merely circulated.
Space is the only thing you cannot buy in the transfer market. Clean data is the same. No subscription package sells you a dataset free of labelling error, because such errors are born of people and process, not of infrastructure.

Tactics are not a diagram on a whiteboard; they are habits repeated across ninety minutes. Data works the same way: its value lies not in one handsome match, but in reliability repeated across hundreds of them.
What I want to see this season is not another imported metric. I want to see a public verification log: every analysis stating its event-data source, its publication date, the definition of each metric used, the sample size, and the places where the author is unsure. Whoever builds that first will win with far fewer resources than whoever buys the most expensive data package.
As for that text file, I still keep it in a separate folder. It does not belong to football. But it is the cheapest reminder I have ever received about how far a wrong label can travel before somebody finally opens the file and reads it.
